MEDIRL: Predicting the Visual Attention of Drivers via Maximum Entropy Deep Inverse Reinforcement Learning
archive: archived pipeline: cataloged verified
Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)
Summary
This paper addresses the challenge of predicting driver visual attention in accident-prone situations, a critical component for improving the safety of vision-based autonomous vehicles. While existing computer vision models often rely on bottom-up saliency mechanisms that prioritize pixel intensity or color, they fail to capture the top-down, goal-directed attention humans use to identify hazards. The authors propose Maximum Entropy Deep Inverse Reinforcement Learning (MEDIRL), a novel formulation that learns task-sensitive reward functions from expert driver eye-fixation patterns. This approach aims to predict fixation locations that maximize rewards by imitating the efficient attention allocation of attentive drivers, particularly in scenarios leading to rear-end collisions. To support this research, the authors introduce EyeCar, a new dataset capturing driver attention in high-density traffic environments prior to accidents. Unlike previous datasets, EyeCar provides point-of-view footage where the ego-vehicle is involved in the collision, offering a more complex and realistic representation of critical driving events. The MEDIRL model formulates fixation selection as a sequential decision process. It constructs a state representation that simulates the human fovea by combining high-resolution information at the current fixation point with low-resolution peripheral vision. The state also integrates spatial cues, including target features (e.g., brake lights), non-target context (e.g., road lanes), relative distance via depth maps, and driving task classifications (lane-keeping, merging, braking). The model uses maximum entropy deep inverse reinforcement learning to recover the underlying reward function from demonstration trajectories, allowing it to handle the stochastic nature of human visual behavior. The authors evaluated MEDIRL on three established benchmarks (DR(eye)VE, BDD-A, DADA-2000) and the new EyeCar dataset. Results demonstrate that MEDIRL achieves state-of-the-art performance, outperforming existing models in correlation coefficient, shuffled Area Under the Curve (s-AUC), and Kullback-Leibler divergence (KLD) metrics. Specifically, MEDIRL effectively predicted attention across various driving tasks, whereas baseline models often exhibited a strong center bias, ignoring relevant target objects. Ablation studies confirmed that incorporating target and non-target spatial cues, driving tasks, and vehicle speed significantly improved prediction accuracy. Furthermore, models trained on EyeCar generalized well to other benchmarks, indicating that EyeCar’s attention distributions are highly informative for critical situations. The significance of this work lies in its contribution to safer autonomous systems by leveraging human-like, goal-directed visual attention mechanisms. By introducing MEDIRL and the EyeCar dataset, the authors provide a robust framework for predicting where drivers look in hazardous scenarios. This enables autonomous vehicles to better identify relevant visual cues and potential risks, addressing a key limitation in current vision-based driving models. The findings suggest that integrating top-down task-specific attention with bottom-up visual features is essential for reliable hazard detection in complex traffic environments.
Key finding
Modeling driver gaze allocation in accident-prone scenes as inverse reinforcement learning over recorded fixations yields state-of-the-art attention prediction across DR(eye)VE, BDD-A, DADA-2000, and the new EyeCar benchmark, outperforming prior bottom-up and supervised attention models.
Methodology
modeling
Sample size: 20
Provenance
The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via publisher_oa on 2026-05-07 (2 acquisition events logged).
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | — | — | — | 1 | 2026-05-07 |
| archive | success | canonical_url | — | — | 2 | 2026-06-03 |
| extract | success | cached | — | — | 2 | 2026-06-10 |
| clean | success | — | — | — | 1 | 2026-06-01 |
| chunk | success | — | — | — | 1 | 2026-06-01 |
| embed | success | — | — | — | 1 | 2026-06-02 |
| enrich | success | — | — | — | 1 | 2026-05-07 |
| promote | success | — | — | — | 3 | 2026-06-06 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 16 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.
- gaze based attention detection
- distraction detection algorithms
- attention allocation
- situational awareness
- visual
- peripheral attention
Information type
What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).
- Methodological Resource: tool software
- Theoretical Contribution: computational model, theory or model